Abstract: Stroke is a basic component of Chinese characters. Stroke semantic segmentation algorithm is often used in the stroke extraction process of font sketches and AI generated font images. Aimed at the low accuracy of stroke segmentation in current methods, a semantic segmentation algorithm for Chinese stroke images is proposed. An efficient channel attention mechanism for stroke class semantic information was introduced to automatically adjust the convolution kernel size to improve the segmentation accuracy. At the same time, skip connections were constructed between the lower and upper samples of each layer to integrate the characterization information such as the stroke shape of the character image with the deep information such as the outline. The algorithm was tested by using Chinese stroke segmentation data set, and compared with FCN, U-Net and SERT segmentation algorithms. The experimental results show that the segmentation accuracy of the algorithm is better than that of traditional methods in different font style data sets.